If you throw a stone, chances are you’ll hit something that’s not real—like a deepfake. Deepfakes are fake videos or audio recordings that use artificial intelligence (AI) to create convincing but false content. The use of generative AI has made it easier to create these deepfakes, and they are becoming more common online. A recent report from Sumsub, a platform that verifies identities, shows that the number of deepfakes has quadrupled from 2023 to 2024. In fact, deepfakes now make up 7% of all online fraud, which includes things like impersonating someone else, taking over accounts, and tricking people into giving away their personal information.
To help fight against deepfakes, Meta (the company behind Facebook and Instagram) has introduced a new tool called Meta Video Seal. This tool adds a nearly invisible watermark to AI-generated videos, which can help identify them as fake. The announcement was made on Thursday, and the tool is available as open source, meaning that other developers can use and build upon it in their own software. This new tool is part of a broader effort by Meta, which also includes other watermarking tools like Watermark Anything and Audio Seal.
Pierre Fernandez, an AI research scientist at Meta, explained in an interview that they created Video Seal to improve the way videos are marked to show that they are AI-generated and to help protect original content.
Read Also: World Bank May Approve $500M Loan For Nigeria Today
While Video Seal is not the first tool to watermark videos, Fernandez believes that many existing tools have limitations. For instance, some don’t hold up well when videos are compressed, which is common when sharing content on social media. Others may not be efficient enough to use widely or might not be designed specifically for videos, as they were adapted from tools made for images.
In addition to adding a watermark, Video Seal can also embed a hidden message in the video. This message can be revealed later to trace the video back to its source. Meta claims that Video Seal is designed to withstand common editing techniques, such as blurring or cropping, as well as popular compression methods used when sharing videos.
However, Fernandez acknowledges that Video Seal has some drawbacks. There is a balance to strike between how noticeable the watermark is and how well it can resist being altered. If a video is heavily compressed or edited, it might damage the watermark or make it impossible to retrieve.
A significant challenge for Video Seal is encouraging developers and businesses to use it, especially those who already have their own solutions in place. To help with this, Meta is launching a public leaderboard called Meta Omni Seal Bench to compare the effectiveness of different watermarking methods. They are also organizing a workshop on watermarking at ICLR, a major AI conference, to foster collaboration between the industry and academic researchers.
Fernandez expressed hope that more AI researchers and developers will start incorporating some form of watermarking into their work. He emphasized that collaboration with the industry and academic community is essential for advancing the field more quickly.
